feat(v3): T17-T18 模型池 + 工具智能体后端

- T17 模型池：PoolStore（local/budget/premium 条目 + architect/worker/agent 角色指派）， /pool CRUD+连通测试+模型探测端点；build_v2_pipeline 池指派优先（测试 override 最后）； V2Stats 新增 by_model 按 token/成本分账 - T18 智能体：OpenAI 兼容工具调用客户端（transport 可注入）+ AgentService （事件落盘 agent_runs/{id}/events.jsonl）+ /agent 提交/status/events/SSE stream + 工作区浏览/读取端点（越界 400）；轮数与 token 双护栏，模型经池 agent 角色或经典回退 - 新增测试 16 项，全量 262 passed（httpx 假注入，不依赖真实模型/key）
2026-09-01 08:51:52 +08:00
parent 6247e053c1
commit 374dc765c2
8 changed files with 1571 additions and 36 deletions
@@ -0,0 +1,170 @@
+"""智能体端点测试：注入脚本化 chat_fn，不依赖真实模型/API key。"""
+import json
+import time
+
+import pytest
+
+pytest.importorskip("fastapi")
+pytest.importorskip("httpx")
+
+from fastapi.testclient import TestClient
+
+import gateway.agent as ag
+import gateway.api as ga
+from gateway.model_pool import PoolStore
+import gateway.model_pool as mp
+
+
+@pytest.fixture()
+def agent_env(tmp_path, monkeypatch):
+    """隔离：池/服务/工作区全部指向临时目录，chat_fn 用脚本替身。"""
+    mp.reset_pool()
+    mp._store = PoolStore(path=tmp_path / "pool.json")
+    ag.reset_agent_service()
+    service = ag.AgentService(run_dir=tmp_path / "agent_runs")
+    ag._service = service
+    # 快照用户真实设置，测试后原样恢复（settings.json 是活文件，不能污染）
+    store = ga.settings_store()
+    snapshot = json.loads(json.dumps(store._data, ensure_ascii=False))
+    # 工作区指向临时目录 + 给经典回退一个假 key（防止 .env 缺失时 400）
+    store.update({"agent": {"workspace_dir": str(tmp_path / "ws")},
+                  "architect": {"api_key": "sk-fake-test"}})
+    ga.rebuild_pipeline()
+
+    script = []
+
+    def set_script(events):
+        script.clear()
+        script.extend(events)
+
+    def fake_chat_factory(acfg):
+        async def chat_fn(messages, tools_spec):
+            if not script:
+                return {"content": "（脚本用尽）好的。", "tool_calls": [], "usage": {}}
+            return script.pop(0)
+        return chat_fn
+
+    monkeypatch.setattr(ga, "build_agent_chat", fake_chat_factory)
+    yield {"service": service, "set_script": set_script, "ws": tmp_path / "ws"}
+
+    store._data = snapshot
+    store.save()
+    mp.reset_pool()
+    ag.reset_agent_service()
+    ga.rebuild_pipeline()
+
+
+@pytest.fixture()
+def client():
+    return TestClient(ga.app)
+
+
+def _wait_done(service, rid, timeout=10.0):
+    t0 = time.time()
+    while time.time() - t0 < timeout:
+        info = service.get(rid)
+        if info and info.state in ("done", "failed"):
+            return info
+        time.sleep(0.05)
+    return service.get(rid)
+
+
+def test_agent_full_flow(agent_env, client):
+    """写文件 -> 最终答复：验证事件、工作区落盘、状态终态。"""
+    agent_env["set_script"]([
+        {"content": None,
+         "tool_calls": [{"id": "c1", "name": "write_file",
+                         "arguments": {"path": "notes.md", "content": "# 笔记"}}],
+         "usage": {"prompt_tokens": 30, "completion_tokens": 6}},
+        {"content": "已创建 notes.md，任务完成。", "tool_calls": [],
+         "usage": {"prompt_tokens": 40, "completion_tokens": 8}},
+    ])
+    r = client.post("/agent", json={"task": "帮我建一个 notes.md"})
+    assert r.status_code == 200
+    rid = r.json()["request_id"]
+    assert r.json()["status"] == "running"
+
+    info = _wait_done(agent_env["service"], rid)
+    assert info.state == "done", info.error
+    assert "notes.md" in info.response
+
+    # 工作区真实落盘
+    assert (agent_env["ws"] / "notes.md").read_text(encoding="utf-8") == "# 笔记"
+
+    # 事件序列
+    events = client.get(f"/agent/{rid}/events").json()
+    kinds = [e["type"] for e in events]
+    assert "tool_call" in kinds and "tool_result" in kinds and "final" in kinds
+    assert events[-1]["reason"] == "answer"
+
+    # 状态端点
+    st = client.get(f"/agent/{rid}/status").json()
+    assert st["state"] == "done"
+    assert st["prompt_tokens"] == 70 and st["completion_tokens"] == 14
+
+    # 工作区浏览端点
+    ls = client.get("/agent/workspace").json()
+    assert ls["ok"] is True
+    assert any(e["name"] == "notes.md" for e in ls["entries"])
+    f = client.get("/agent/file", params={"path": "notes.md"}).json()
+    assert f["content"] == "# 笔记"
+
+
+def test_agent_jail_via_api(agent_env, client):
+    """工具结果为 ok=False（越界被拒），循环仍能继续到最终答复。"""
+    agent_env["set_script"]([
+        {"content": None,
+         "tool_calls": [{"id": "c1", "name": "read_file",
+                         "arguments": {"path": "../../secret.txt"}}],
+         "usage": {"prompt_tokens": 10, "completion_tokens": 2}},
+        {"content": "越界访问被拒绝。", "tool_calls": [], "usage": {}},
+    ])
+    r = client.post("/agent", json={"task": "读一下上级目录"})
+    rid = r.json()["request_id"]
+    info = _wait_done(agent_env["service"], rid)
+    assert info.state == "done"
+    events = agent_env["service"].read_events(rid)
+    tool_result = next(e for e in events if e["type"] == "tool_result")
+    assert tool_result["ok"] is False
+
+    # 文件读取 API 直接越界 -> 404/400
+    r2 = client.get("/agent/file", params={"path": "../../x.txt"})
+    assert r2.status_code in (400, 404)
+
+
+def test_agent_model_from_pool(agent_env, client, monkeypatch):
+    """池 agent 角色（或显式 pool_id）应被采用；mock 池模型拒绝。"""
+    from gateway.agent import OpenAICompatChat
+    captured = {}
+    real_factory = None
+
+    # 先放一个 openai 池条目并指派 agent 角色
+    client.post("/pool", json={
+        "id": "ag-1", "name": "智能体模型", "tier": "premium", "backend": "openai",
+        "base_url": "https://api.example.com", "model": "big-model-x",
+        "api_key": "sk-abc1234567", "enabled": True,
+    })
+    client.put("/pool/roles", json={"agent": "ag-1"})
+
+    # /agent 不带 pool_id -> 用池 agent 角色
+    r = client.post("/agent", json={"task": "hi"})
+    assert r.status_code == 200
+    assert r.json()["model"] == "big-model-x"
+
+    # mock 条目 -> 400
+    client.post("/pool", json={
+        "id": "mk-1", "name": "mock", "tier": "local", "backend": "mock",
+        "model": "mock", "enabled": True,
+    })
+    r2 = client.post("/agent", json={"task": "hi", "pool_id": "mk-1"})
+    assert r2.status_code == 400
+
+
+def test_agent_task_validation(agent_env, client):
+    assert client.post("/agent", json={"task": ""}).status_code == 400
+    assert client.post("/agent", json={}).status_code == 400
+
+
+def test_agent_404(agent_env, client):
+    assert client.get("/agent/ghost/status").status_code == 404
+    assert client.get("/agent/ghost/events").json() == []
@@ -0,0 +1,171 @@
+"""模型池（PoolStore）测试：条目校验/CRUD/角色指派/管线解析/成本分账。"""
+import pytest
+
+pytest.importorskip("fastapi")
+
+from fastapi.testclient import TestClient
+
+import gateway.api as ga
+import gateway.model_pool as mp
+from gateway.model_pool import PoolStore, compute_cost, entry_to_architect_cfg, entry_to_worker_cfg
+
+
+@pytest.fixture()
+def pool(tmp_path):
+    """独立文件的全局池（不污染 config/model_pool.json）。"""
+    mp.reset_pool()
+    store = PoolStore(path=tmp_path / "model_pool.json")
+    mp._store = store
+    yield store
+    mp.reset_pool()
+    ga.rebuild_pipeline()
+
+
+@pytest.fixture()
+def client():
+    return TestClient(ga.app)
+
+
+def _entry(**over):
+    base = {
+        "id": "prem-1", "name": "旗舰模型", "tier": "premium", "backend": "openai",
+        "base_url": "https://api.deepseek.com", "model": "deepseek-v4-pro",
+        "api_key": "sk-test-1234567890", "price_in": 1.0, "price_out": 2.0,
+        "enabled": True,
+    }
+    base.update(over)
+    return base
+
+
+# ---------------- PoolStore 单元 ----------------
+
+def test_pool_upsert_and_mask(pool):
+    masked = pool.upsert(_entry())
+    assert masked["api_key_set"] is True
+    assert "sk-test" not in masked["api_key"]          # 明文不打回
+    data = pool.list()
+    assert data["entries"][0]["model"] == "deepseek-v4-pro"
+    assert data["entries"][0]["api_key_set"] is True
+
+
+def test_pool_upsert_keeps_key_when_blank(pool):
+    pool.upsert(_entry())
+    pool.upsert(_entry(api_key=""))                    # 前端不回传明文 -> 保留
+    assert pool.get("prem-1")["api_key"] == "sk-test-1234567890"
+
+
+def test_pool_validation(pool):
+    with pytest.raises(ValueError):
+        pool.upsert(_entry(tier="超豪华"))
+    with pytest.raises(ValueError):
+        pool.upsert(_entry(backend="magic"))
+    with pytest.raises(ValueError):
+        pool.upsert(_entry(backend="openai", base_url=""))   # 非 mock 缺端点
+    with pytest.raises(ValueError):
+        pool.upsert(_entry(hack="x"))                        # 未知字段
+    with pytest.raises(ValueError):
+        pool.upsert(_entry(price_in=-1))
+
+
+def test_pool_roles_and_resolve(pool):
+    pool.upsert(_entry())
+    pool.upsert(_entry(id="local-1", tier="local", backend="llama_server",
+                       base_url="http://127.0.0.1:8901/v1", model="qwen3.5-4b",
+                       price_in=0, price_out=0))
+    assert pool.resolve("architect") is None              # 未指派
+    pool.set_roles({"architect": "prem-1", "worker": "local-1"})
+    assert pool.resolve("architect")["id"] == "prem-1"
+    assert pool.resolve("worker")["id"] == "local-1"
+    assert pool.resolve("agent") is None
+    # 指派不存在的条目
+    with pytest.raises(ValueError):
+        pool.set_roles({"agent": "ghost"})
+    # 删除条目 -> 角色自动清空
+    pool.delete("prem-1")
+    assert pool.resolve("architect") is None
+
+
+def test_pool_disabled_entry_not_resolved(pool):
+    pool.upsert(_entry(enabled=False))
+    pool.set_roles({"architect": "prem-1"})
+    assert pool.resolve("architect") is None              # 禁用 -> 回退经典设置
+
+
+def test_entry_cfg_mapping(pool):
+    e = pool.get("prem-1") or _entry()
+    acfg = entry_to_architect_cfg(_entry())
+    assert acfg["model"] == "deepseek-v4-pro"
+    assert acfg["api_key"] == "sk-test-1234567890"
+    wcfg = entry_to_worker_cfg(_entry())
+    assert wcfg["backend"] == "openai"
+
+
+def test_compute_cost():
+    e = {"price_in": 1.0, "price_out": 2.0}
+    assert compute_cost(e, 1_000_000, 500_000) == pytest.approx(2.0)
+    assert compute_cost({"price_in": 0, "price_out": 0}, 999, 999) == 0.0
+
+
+# ---------------- API 端点 ----------------
+
+def test_pool_api_crud(pool, client):
+    r = client.get("/pool")
+    assert r.status_code == 200
+    assert r.json()["roles"]["architect"] == ""
+    r2 = client.post("/pool", json=_entry())
+    assert r2.status_code == 200
+    assert len(r2.json()["entries"]) == 1
+    # 非法条目 -> 400
+    r3 = client.post("/pool", json=_entry(tier="bad"))
+    assert r3.status_code == 400
+    # 角色指派
+    r4 = client.put("/pool/roles", json={"architect": "prem-1"})
+    assert r4.status_code == 200
+    assert r4.json()["roles"]["architect"] == "prem-1"
+    # 删除
+    r5 = client.delete("/pool/prem-1")
+    assert r5.status_code == 200
+    assert r5.json()["roles"]["architect"] == ""
+
+
+def test_build_pipeline_uses_pool(pool, monkeypatch):
+    """池指派应覆盖经典设置，测试 override 最后生效。"""
+    pool.upsert(_entry())
+    pool.set_roles({"architect": "prem-1"})
+    ga.rebuild_pipeline()
+
+    captured = {}
+
+    def fake_build_architect(cfg):
+        captured["architect"] = dict(cfg)
+        from router_system.architect import ArchitectClient
+        return ArchitectClient(model=cfg.get("model", "m"), api_key="k")
+
+    monkeypatch.setattr(ga, "build_architect", fake_build_architect)
+    pipe = ga.build_v2_pipeline(worker_cfg_override={"backend": "mock"})
+    assert pipe is not None
+    assert captured["architect"]["model"] == "deepseek-v4-pro"   # 池条目生效
+    ga.rebuild_pipeline()
+
+
+def test_v2stats_by_model():
+    from router_system.v2stats import V2Stats
+
+    class R:
+        request_id = "x"
+        fast_path = False
+        status = "done"
+        rounds_used = 1
+        api_input_tokens = 1000
+        api_output_tokens = 500
+        cost_est = 0.002
+        model_used = "deepseek-v4-pro"
+        route = []
+
+    s = V2Stats()
+    s.record(R())
+    summary = s.summary()
+    bucket = summary["by_model"]["deepseek-v4-pro"]
+    assert bucket["requests"] == 1
+    assert bucket["input_tokens"] == 1000
+    assert bucket["cost_est_usd"] == pytest.approx(0.002)